METHANE-PLUME SEGMENTATION FROM HYPERSPECTRAL SATELLITE IMAGERY VIA MULTIMODAL DEEP LEARNING

Accurate detection and segmentation of methane plumes from satellite imagery are essential for climate monitoring, yet these tasks are hindered by the low signal-to-noise ratio and spatial variability of gas emissions. The authors address the challenge of creating a scalable, efficient, and accurate deep learning model that can effectively fuse spectral methane data with contextual RGB imagery.

The researchers propose a multimodal architecture that processes RGB imagery and methane enhancement maps through separate encoders. The core innovation is the Feature-Guided Methane Enhancement (FGME) mechanism, which injects physically meaningful methane cues into the RGB feature maps at multiple semantic scales. This allows the model to leverage both the spectral evidence of methane and the spatial context provided by the RGB data. The resulting fused features are processed by a SegFormer-based decoder, followed by a refinement head to sharpen plume boundaries.

Evaluated on the MPDataset, the proposed model outperforms existing state-of-the-art methods, including MPSUNet, across all key metrics: Mean Intersection over Union (MIoU), Mean Precision, and Recall. Notably, the model achieves these improvements while requiring approximately 80% fewer floating-point operations (FLOPs) than previous high-performing architectures. Qualitative analysis demonstrates that the model produces more spatially coherent and physically consistent plume masks, effectively reducing the "hallucination" of spurious plume fragments often seen in other models.

This work provides a robust and computationally efficient solution for large-scale methane monitoring. By demonstrating that multimodal fusion can improve segmentation precision while lowering the computational burden, the study offers a practical pathway for integrating high-resolution satellite data into real-world environmental monitoring systems.